Periodic Reporting for period 2 - PHRASE (Personalised Health cognitive assistance for RehAbilitation SystEm)
Reporting period: 2023-04-01 to 2025-07-31
Stroke often leaves lasting motor and cognitive difficulties. Access to high-quality rehabilitation is uneven and benefits can diminish after discharge, so health systems need safe, home-ready digital tools that extend therapy and give clinicians objective, timely insight. PHRASE was designed in this context to connect data capture, interpretation and intervention in one pathway for stroke rehabilitation.
The project’s overall objective was to deliver an end-to-end prototype that clinicians can prescribe and monitor, and that patients can use at home under governed European data practices. PHRASE integrates three pillars: EBRAINS for federated, FAIR data and knowledge-graph services; a GDPR-aligned Virtual Research Environment for secure analytics; and the RGS suite (clinic and home apps plus dashboards) for assessment, training and review. On top, AI components estimate status from day-to-day interactions, forecast recovery trajectories, and adapt training difficulty to keep patients engaged.
Pathway to impact: (1) capture and harmonise clinic and home data without moving sensitive records; (2) turn signals into actionable, interpretable recommendations; (3) embed them in tools patients and therapists already use; and (4) generate evidence through feasibility and randomised evaluation to inform adoption and future CE-marking paths. At scale, this can raise therapy “dose” at home, target clinic resources more efficiently, and reduce avoidable disability across ageing populations.
The project’s overall objective was to deliver an end-to-end prototype that clinicians can prescribe and monitor, and that patients can use at home under governed European data practices. PHRASE integrates three pillars: EBRAINS for federated, FAIR data and knowledge-graph services; a GDPR-aligned Virtual Research Environment for secure analytics; and the RGS suite (clinic and home apps plus dashboards) for assessment, training and review. On top, AI components estimate status from day-to-day interactions, forecast recovery trajectories, and adapt training difficulty to keep patients engaged.
Pathway to impact: (1) capture and harmonise clinic and home data without moving sensitive records; (2) turn signals into actionable, interpretable recommendations; (3) embed them in tools patients and therapists already use; and (4) generate evidence through feasibility and randomised evaluation to inform adoption and future CE-marking paths. At scale, this can raise therapy “dose” at home, target clinic resources more efficiently, and reduce avoidable disability across ageing populations.
PHRASE delivered a complete, GDPR-aligned pipeline that links data capture, analysis, and personalised stroke rehabilitation. The core infrastructure is a hospital-operated virtual research environment, integrated with a secure research cloud and a knowledge graph, so sensitive data are processed in isolated, auditable workspaces while non-sensitive metadata remain findable. This backbone supported all clinical data flows.
Patient and clinician apps were upgraded with clearer onboarding, short in-app tutorials, a virtual coach with reminders, and structured at-home assessments of motor and cognitive function. Wearable-based arm-use monitoring and an adaptive difficulty engine were hardened through cross-device testing to improve reliability.
AI components moved from prototypes to evaluated modules embedded in the pipeline. A “digital diagnosis” estimates standard clinical scales from day-to-day interaction data and reports confidence; a prognosis model predicts recovery trajectories on widely used stroke scales; and training is personalised automatically. Internal evaluations on harmonised datasets showed strong agreement with clinical measures and accurate class predictions, and outputs are generated server-side and shown in clinician dashboards for remote review.
Clinical work progressed from feasibility to a randomised controlled trial. The feasibility study enrolled 88 participants across five centres and reported good usability (System Usability Scale 70.21). The trial was registered and approved, sites were activated, electronic case-report forms configured, and enrolment began; confirmatory effectiveness and agreement analyses will follow completion as planned.
In sum, PHRASE delivered a secure analytics backbone, a hardened telerehabilitation suite, uncertainty-aware diagnosis and prognosis integrated into clinician tools, and feasibility evidence with a live pathway to confirmatory trial outcomes.
Patient and clinician apps were upgraded with clearer onboarding, short in-app tutorials, a virtual coach with reminders, and structured at-home assessments of motor and cognitive function. Wearable-based arm-use monitoring and an adaptive difficulty engine were hardened through cross-device testing to improve reliability.
AI components moved from prototypes to evaluated modules embedded in the pipeline. A “digital diagnosis” estimates standard clinical scales from day-to-day interaction data and reports confidence; a prognosis model predicts recovery trajectories on widely used stroke scales; and training is personalised automatically. Internal evaluations on harmonised datasets showed strong agreement with clinical measures and accurate class predictions, and outputs are generated server-side and shown in clinician dashboards for remote review.
Clinical work progressed from feasibility to a randomised controlled trial. The feasibility study enrolled 88 participants across five centres and reported good usability (System Usability Scale 70.21). The trial was registered and approved, sites were activated, electronic case-report forms configured, and enrolment began; confirmatory effectiveness and agreement analyses will follow completion as planned.
In sum, PHRASE delivered a secure analytics backbone, a hardened telerehabilitation suite, uncertainty-aware diagnosis and prognosis integrated into clinician tools, and feasibility evidence with a live pathway to confirmatory trial outcomes.
PHRASE moved beyond the state of the art by operating a hospital-grade Virtual Research Environment federated with EBRAINS so sensitive, multi-site neurorehabilitation data could be processed under GDPR with provenance and Knowledge Graph discoverability, avoiding raw data transfers while keeping metadata findable for reuse. It delivered uncertainty-aware digital diagnosis that converts in-app kinematics and task performance into weekly clinical-scale estimates for remote review, using a Bayesian committee to fuse sources with confidence intervals exposed to clinicians. Prognosis advanced through a longitudinal mixture model retrained on harmonised datasets, extending from FM-UE to ARAT and achieving internal performance around ~81% correct recovery class with two assessments (~89% with five), with ~5-point MAE at 24 weeks; figures will be re-estimated on the RCT dataset. The Personalised Ability Manifold for training was strengthened with calibrated difficulty ranges and a binomial-likelihood, MCMC-based inference to quantify uncertainty and support both rapid personalisation and diagnosis-ready biomarkers. These capabilities were embedded in a hardened telerehabilitation suite (RGSapp, MIMS, RGSwear) validated through device-matrix tests and a multi-site feasibility study reporting SUS 70.21 above the canonical benchmark, underpinning the launch of a prospectively registered RCT with activated sites and compliant eCRF operations.
Indicative impacts include increasing at-home therapy dose, enabling objective between-visit monitoring, and focusing therapist time where it matters, while giving health systems governed, reusable data assets for research and service improvement. For further uptake and success, the immediate needs are completion and publication of confirmatory RCT outcomes with re-estimation of diagnostic and prognostic performance; regulatory pathway work towards CE-marking with documentation of AI transparency; interoperability with site systems to reduce adoption friction; and a scale-out plan that couples business model, reimbursement arguments and an appropriate IPR posture so algorithms are protected while interfaces and metadata remain open for ecosystem integration.
Indicative impacts include increasing at-home therapy dose, enabling objective between-visit monitoring, and focusing therapist time where it matters, while giving health systems governed, reusable data assets for research and service improvement. For further uptake and success, the immediate needs are completion and publication of confirmatory RCT outcomes with re-estimation of diagnostic and prognostic performance; regulatory pathway work towards CE-marking with documentation of AI transparency; interoperability with site systems to reduce adoption friction; and a scale-out plan that couples business model, reimbursement arguments and an appropriate IPR posture so algorithms are protected while interfaces and metadata remain open for ecosystem integration.